Search

Thin leaves, fine stems, self-occlusion, non-rigid and slowly changing structures make plants difficult for three-dimensional (3D) scanning and reconstruction - two critical steps in automated visual phenotyping. Many current solutions such as laser scanning, structured light, and multiview stereo can struggle to acquire usable 3D models because of limitations in scanning resolution and calibration accuracy. In response, we have developed a fast, low-cost, 3D scanning platform to image plants on a rotating stage with two tilting DSLR cameras centred on the plant. This uses new methods of camera calibration and background removal to achieve high-accuracy 3D reconstruction. We assessed the system's accuracy using a 3D visual hull reconstruction algorithm applied on 2 plastic models of dicotyledonous plants, 2 sorghum plants and 2 wheat plants across different sets of tilt angles. Scan times ranged from 3 minutes (to capture 72 images using 2 tilt angles), to 30 minutes (to capture 360 images using 10 tilt angles). The leaf lengths, widths, areas and perimeters of the plastic models were measured manually and compared to measurements from the scanning system: results were within 3-4\% of each other. The 3D reconstructions obtained with the scanning system show excellent geometric agreement with all six plant specimens, even plants with thin leaves and fine stems.

Hyperspectral images typically contain hundreds of spectral bands, which is one to two orders of magnitude larger than the number of bands in multispectral images. This greater volume of spectral information could lead to discoveries that are not possible with multispectral images; however, overcoming the complexity of the additional information is a computational challenge. Here, we present a solution that uses feature selection, logarithmic nearest neighbor classification and neighborhood spatial analysis to classify the land use of multiple hyperspectral images. Empirical analysis shows that our solution is as accurate as other much more complex approaches and it is orders-of-magnitude more efficient. This ascertains that our solution is scalable to larger datasets while maintaining high accuracy.

A method for segmenting HEp-2 cells in Indirect Immuno Fluorescence microscope(IIF) images is implemented and evaluated. Challenges to accurate segmentation include the complexity of the data acquired at multiple wavelengths, overlapping cells, and variations in staining. Level set methods via geometric active contours are used to solve this problem. Level set methods use morphological operations to estimate an initial cell boundary and are fully automated. Geometric active contours are able to adapt to the curve topology of the cell boundary. Segmentation performance is evaluated using six indices: boundary displacement error, global consistency error, variation of information, Jaccard distance error, rand index and F-index .

The forced landing problem has become one of the main impediments to UAV's entering civilian airspace. Unfortunately there is no robust forced landing site detection system that will reliably detect a safe landing site. One of the main reasons for this is the difficulty in considering the various classes of surface, to determine whether they are safe or not. We propose a robust UAV landing site detection system using mid-level discriminative patches. The training and tuning process uses a dataset containing 1600 randomly selected Google map images with weak labels. We then show how the output from multiple mid-level discriminative patch detectors can be combined to indicate the level or danger for a given region. The proposed technique reliably detects safe landing areas in UAV imagery, and achieves improved performance over the state-of-the art. The proposed system outperforms the baseline system by 29.4% for completeness and 33.9% for correctness, and is invariant to the changes of illumination, sharpness and resolution of images.

When processing video, it is normally assumed that cameras are vertically oriented such that people appear upright, which helps simplify subsequent processing such as person detection. In real situations, due to the need to provide maximum coverage of the viewing space, cameras are usually placed with arbitrary orientations so the apparent vertical axis of the videos captured may not correspond to the true vertical direction of the captured scene. To rectify this situation, we propose a classification-based system, which normalizes the video compensating for the camera orientation. We demonstrate the performance of the system for outdoor sports video. Our system works as follows: From an arbitrary set of sports videos, we first automatically create a training/testing image dataset, in which players have various orientations. Our classifier is a stacked autoencoder connected to a softmax output layer, which is trained using this dataset for estimating the orientation of players. The orientation of an input video is normalized according to the orientations of player patches, whose angles of orientation are estimated by the above trained classifier. The experiments conducted on hockey field video dataset show that the proposed system is able to estimate the true vertical axis of an input video accurately.

In this paper the problem of complex event detection is addressed. Existing event detection methods are limited to features that are extracted from local spatial or spatio-temporal patches from the videos. However, this makes the model more vulnerable to the events that have similar concepts with different actions e.g. "Open drawer" and "Open cupboard". Furthermore, current methods typically assume that events have already been segmented from the video stream, and do not generalize well to events with unknown starting and ending locations. In this work, in order to address the aforementioned limitations we present a novel model based on the combination of semantic and temporal features extracted from video frames. We train a max-margin classifier on top of the extracted features in an adaptive framework that is able to detect the events with unknown starting and ending locations. Our model is based on the Bidirectional Region Neural Network and large margin Structural Output SVM. The generality of our model allows it to be simply applied on different labeled and unlabeled datasets.We finally test our algorithm on three challenging datasets: "UCF 101-Action Recognition", "MPII Cooking Activities" and "Hollywood"; and we report the state-of-the-art performance.

A new method for segmenting white blood cells nuclei in microscopic images is presented. Challenges to accurate segmentation include intra-class variation of the nuclei cell boundaries, non-uniform illumination, and changes in the cell topology due to its orientation and stage of maturity. In this research, level set methods and geometric active contours are used to segment the nucleus of white blood cells from the cytoplasm and the cell wall. Level set methods use morphological operations to estimate an initial cell boundary and are fully automated. Geometric active contours are less computationally complex and adapt better to the curve topology of the cell boundary than parametric active contours, which have been previously used for white blood cell segmentation. Segmentation performance is compared with other segmentation methods using the Berkeley benchmark database and the proposed method is shown to be superior using various indices.

This paper addresses the challenge of introducing a Bayes rules to measure flood probability from multispectral data. Machine learning classifiers were applied to map the extent of flood inundation from multispectral remote sensing imagery. The paper applies Extended Support Vector Machine classifier along with linear spectral unmixing to obtain the classification output. K-means clustering is applied on pre and post flood images to select SVM training samples from clustering outcome of the most informative spectral band. Experiments were conducted by dividing training and testing samples into two groups. The efficiency of classifier was enhanced by introducing the Bayesian probability measure and performance was assessed by using precision and recall metrics on the pre and post Bayesian flood probability estimation. It has been observed that for some test cases in this study there was a substantial improvement in precision-recall curve with high precision values and low recall rate. The optimal flood probability threshold value has also been easily calculated by calculating and analising iteratively precision and recall.

Person re-identification is particularly challenging due to significant appearance changes across separate camera views. In order to re-identify people, a representative human signature should effectively handle differences in illumination, pose and camera parameters. While general appearance-based methods are modelled in Euclidean spaces, it has been argued that some applications in image and video analysis are better modelled via non-Euclidean manifold geometry. To this end, recent approaches represent images as covariance matrices, and interpret such matrices as points on Riemannian manifolds. As direct classification on such manifolds can be difficult, in this paper we propose to represent each manifold point as a vector of similarities to class representers, via a recently introduced form of Bregman matrix divergence known as the Stein divergence. This is followed by using a discriminative mapping of similarity vectors for final classification. The use of similarity vectors is in contrast to the traditional approach of embedding manifolds into tangent spaces, which can suffer from representing the manifold structure inaccurately. Comparative evaluations on benchmark ETHZ and iLIDS datasets for the person re-identification task show that the proposed approach obtains better performance than recent techniques such as Histogram Plus Epitome, Partial Least Squares, and Symmetry-Driven Accumulation of Local Features.

We present a novel approach for multi-object detection in aerial videos based on tracking. The proposed method mainly involves three steps. Firstly, the spatial-temporal saliency is employed to detect moving objects. Secondly, the detected objects are tracked by mean shift in the subsequent frames. Finally, the saliency results are fused with the weight map generated by tracking to get refined detection results, and in turn the modified detection results are used to update the tracking models. The proposed algorithm is evaluated on VIVID aerial videos, and the results show that our approach can reliably detect moving objects even in challenging situations. Meanwhile, the proposed method can process videos in real time, without the effect of time delay.

Time-Frequency analysis based on the Wigner-Ville Distribution (WVD) is shown to be optimal for a class of signals where the variation of instantaneous frequency is the dominant characteristic. Spectral resolution and instantaneous frequency tracking is substantially improved by using a Modified WVD (MWVD) based on an Autoregressive spectral estimator. Enhanced signal-to-noise ratio may be achieved by using 2D windowing in the Time-Frequency domain. The WVD provides a tool for deriving descriptors of signals which highlight their FM characteristics. These descriptors may be used for pattern recognition and data clustering using the methods presented in this paper., This paper presents a modified Wigner-Ville Distribution (MWVD) using Auto-regressive spectral analysis to improve the spectral resolution; it also presents descriptors which highlight a signal's FM characteristics and a data clustering algorithm for classification. (Additional details can be found in the comprehensive book on Time-Frequency Signal Analysis and Processing (see http://www.elsevier.com/locate/isbn/0080443354). In addition, the most recent upgrade of the original software package that calculates Time-Frequency Distributions and Instantaneous Frequency estimators can be downloaded from the web site: www.time-frequency.net. This was the first software developed in the field, and it was first released publicly in 1987 at the 1st ISSPA conference held in Brisbane, Australia, and then continuously updated).

Soft biometrics are characteristics that can be used to describe, but not uniquely identify an individual. These include traits such as height, weight, gender, hair, skin and clothing colour. Unlike traditional biometrics (i.e. face, voice) which require cooperation from the subject, soft biometrics can be acquired by surveillance cameras at range without any user cooperation. Whilst these traits cannot provide robust authentication, they can be used to provide coarse authentication or identification at long range, locate a subject who has been previously seen or who matches a description, as well as aid in object tracking. In this paper we propose three part (head, torso, legs) height and colour soft biometric models, and demonstrate their verification performance on a subset of the PETS 2006 database. We show that these models, whilst not as accurate as traditional biometrics, can still achieve acceptable rates of accuracy in situations where traditional biometrics cannot be applied.

Light Detection and Ranging (LIDAR) has great potential to assist vegetation management in power line corridors by providing more accurate geometric information of the power line assets and vegetation along the corridors. However, the development of algorithms for the automatic processing of LIDAR point cloud data, in particular for feature extraction and classification of raw point cloud data, is in still in its infancy. In this paper, we take advantage of LIDAR intensity and try to classify ground and non-ground points by statistically analyzing the skewness and kurtosis of the intensity data. Moreover, the Hough transform is employed to detected power lines from the filtered object points. The experimental results show the effectiveness of our methods and indicate that better results were obtained by using LIDAR intensity data than elevation data.

In public venues, crowd size is a key indicator of crowd safety and stability. Crowding levels can be detected using holistic image features, however this requires a large amount of training data to capture the wide variations in crowd distribution. If a crowd counting algorithm is to be deployed across a large number of cameras, such a large and burdensome training requirement is far from ideal. In this paper we propose an approach that uses local features to count the number of people in each foreground blob segment, so that the total crowd estimate is the sum of the group sizes. This results in an approach that is scalable to crowd volumes not seen in the training data, and can be trained on a very small data set. As a local approach is used, the proposed algorithm can easily be used to estimate crowd density throughout different regions of the scene and be used in a multi-camera environment. A unique localised approach to ground truth annotation reduces the required training data is also presented, as a localised approach to crowd counting has different training requirements to a holistic one. Testing on a large pedestrian database compares the proposed technique to existing holistic techniques and demonstrates improved accuracy, and superior performance when test conditions are unseen in the training set, or a minimal training set is used.

The relationship between multiple cameras viewing the same scene may be discovered automatically by finding corresponding points in the two views and then solving for the camera geometry. In camera networks with sparsely placed cameras, low resolution cameras or in scenes with few distinguishable features it may be difficult to find a sufficient number of reliable correspondences from which to compute geometry. This paper presents a method for extracting a larger number of correspondences from an initial set of putative correspondences without any knowledge of the scene or camera geometry. The method may be used to increase the number of correspondences and make geometry computations possible in cases where existing methods have produced insufficient correspondences.

Person tracking systems are dependent on being able to locate a person accurately across a series of frames. Optical flow can be used to segment a moving object from a scene, provided the expected velocity of the moving object is known; but successful detection also relies on being able segment the background. A problem with existing optical flow techniques is that they don’t discriminate the foreground from the background, and so often detect motion (and thus the object) in the background. To overcome this problem, we propose a new optical flow technique, that is based upon an adaptive background segmentation technique, which only determines optical flow in regions of motion. This technique has been developed with a view to being used in surveillance systems, and our testing shows that for this application it is more effective than other standard optical flow techniques.

Video surveillance systems are becoming an indispensable tool in today's environment, particularly for security related applications. Surveillance footage is often routinely used to identify faces of criminals "caught in the act" or for tracking individuals in a crowded environment. Most face images captured with these systems however, are small and coarse, making it extremely difficult to identify an individual through human observation or via automatic face recognition systems. Super-resolution (SR) is a technique that can overcome this limitation by combining complimentary information from several frames of a video sequence to produce high resolution images of a subject. A problem that plagues many existing SR systems is that they can only deal with simple, rigid inter-frame transformations, thus performing poorly with face images as faces are non-planar, non-rigid, non-lambertian and can self-occlude. This paper presents a SR system to overcome these limitations by using a robust optical flow technique. An investigation into the quality of the super-resolved images and their dependency on the number of video sequence frames used in the reconstruction is undertaken. Different fusion techniques are also investigated and experiments are conducted over two image sequences. Results show significant improvement of the image quality and resolution over the original low resolution sequences.

As new genome sequencing initiatives are completed, one of the next great challenges of cell biology is the atomic resolution structure determination of the enormous number of proteins they encode. Single particle analysis is a technique which produces 3D structures by computationally aligning high resolution electron microscope images of individual, randomly oriented molecules. One of the limiting factors in producing a high resolution 3D reconstruction is obtaining a large enough representative dataset (~100,000 particles). Traditionally particles have been picked manually but this is a slow and labour intensive process. This paper describes two automatic particle picking algorithms, based on correlation and edge detection, which have been shown to be capable of quickly selecting a large number of particles in micrographs. Currently circular and rectangular particles are able to be picked.

Much recent research attention has focused on providing some form of visually meaningful information to blind people through electrical stimulation of a component of the visual system. Current technology limits the number of perceived points of light (phosphenes) that can be provided to a user and methods are required to optimize the amount of presented information. This paper describes a PDA based artificial human vision simulator, and proposes a method for alerting a user of possible looming obstacles. Experimental results indicate that obstacle alerts can be successfully provided, however with the current simulator components, high-quality lighting and accurate image segmentation is critical for reducing the number of false alerts.

Uniform sensor management and abstraction across different robot platforms is a difficult task due to the sheer diversity of sensing devices. However, because these sensors can be grouped into categories that in essence provide the same information, we can capture their similarities and create abstractions. An example would be distance data measured by an assortment of range sensors, or alternatively extracted from a camera using image processing. This paper describes how using software components it is possible to uniformly construct high-level abstractions of sensor information across various robots in a way to support the portability of common code that uses these abstractions (e.g. obstacle avoidance, wall following). We demonstrate our abstractions on a number of robots using different configurations of range sensors and cameras.

Kernel methods such as kernel principal component analysis and support vector machines have become powerful tools for pattern recognition and computer vision. Unfortunately the high computational cost of kernel methods is a limiting factor for real-time classification tasks when running on the CPU of a standard PC. Over the last few years, commodity Graphics Processing Units (GPU) have evolved from fixed graphics pipeline processors into more flexible and powerful data-parallel processors. These stream processors are capable of sustaining computation rates of greater than ten times that of a single CPU. GPUs are inexpensive and are becoming ubiquitous (desktops, laptops, PDAs, cell phones). In this paper, we present a face recognition system based on kernel methods running on the GPU. This GPU implementation is twenty eight times faster than the same optimized application running on the CPU.

Coronary heart disease was the single largest cause of sudden death in Australia in 2002. Computed tomography angiography (CTA) provides high resolution, high contrast images of the thoracic cavity, and as such has emerged as the imaging modality of choice for diagnosing and planning treatment for coronary heart disease. However, radiologists and cardiac surgeons require tools to easily identify possible stenosis (narrowed or constricted coronary vessels) in such CTA datasets. We present a method which allows users to interactively visualise a specific three-dimensional region of interest (ROI). In our example, segmentation methods are applied to isolate the coronary vessels, which in turn are visually enhanced using various perceptual cues. The segmentation is achieved using a combination of thresholding, region-growing, and morphological operations. The perceptual enhancement is realized by fusing direct volume rendered images using weighting factors determined by the segmented regions. The user can allow for the easy dissemination of relevant information by adjusting 'transfer functions' to control the degree of ROI enhancement. This approach requires only roughly segmented regions of interest, and allows for the 3D visualisation of calcifications within vessels. This proposed method has significant potential for helping to facilitate the efficient treatment for coronary heart disease. Furthermore, it can be implemented at interactive framerates on comparatively cheap, desktop comuting hardware making it readily accessible.

Direct volume rendering (DVR) provides medical users with insight into datasets by creating a 3-D representation from a set of 2-D image slices (such as CT or MRI). This visualisation technique has been used to aid various medi-cal diagnostic and therapy planning tasks. Volume render-ing has recently become faster and more affordable with the advent of 3-D texture-mapping on commodity graphics hardware. Current implementations of the DVR algorithm on such hardware allow users to classify sample points (known as "voxels") using 2-D transfer functions (func-tions based on sample intensity and sample intensity gradi-ent magnitude). However, such 2-D transfer functions in-herently ignore spatial information. We present a novel modification to 3-D texture-based volume rendering allow-ing users to classify fuzzy-segmented, overlapping regions with independent 2-D transfer functions. This modification improves direct volume rendering by allowing for more sophisticated classification using spatial information.

Mobile robots operating in unknown environments need to build maps. To do so they must have an exploration algorithm to plan a path. This algorithm should guarantee that the whole of the environment, or at least some designated area, will be mapped. The path should also be optimal in some sense and not simply a "random walk" which is clearly inefficient. When multiple robots are involved, the algorithm also needs to take advantage of the fact that the robots can share the task. In this paper we discuss a modification to the well-known distance transform that satisfies these requirements.

Medical images in digital form must be stored in a secure way to preserve stringent image quality standards and prevent unauthorised disclosure of patient data. This paper proposes a multiple watermarking method to serve these purposes. A multiple watermark consists of an annotation part and a fragile part. Encrypted patient data can be embedded in an annotation watermark, and tampering can be detected using a fragile watermark. The embedded patient data not only save storage space, it also offers privacy and security. We also evaluate the images’ visual quality after watermark embedding and the effectiveness of locating tampered regions.

This paper is a result of a fruitful cooperation between the computer science and the dental diagnosis experiences. The study presents a new approach of applying computer algorithms to radiographic images of dental implantation used for bone regeneration. We focus here only on the contribution of the computer assistance to the clinical research as the periodontal therapy is beyond the scope of this paper. The proposed system is based on a pattern recognition approach, directed to recognize density changes in the intra-bony affected areas of patients. It comprises different modules with new algorithms specially designed to treat the patients’ radiographic images more accurately. The system includes digitizing, detecting the complicated region of interest (ROI), defining reference area to correct any projection discrepancy of the follow up images, and finally to extract the distinguishing features of the ROI as a basis for determining the rate of new bone density accumulation. This study is applied to two typical dental cases for a patient who received two different operations. The results are very encouraging and more accurate than traditional techniques reported before.